Case Study: The State-Preparation Problem
Executive Summary
Quantum phase estimation is often described as the algorithm that will deliver exact molecular ground-state energies and thereby transform chemistry and materials science. The description is accurate about what QPE computes and silent about a prerequisite that frequently dominates the cost: you must first prepare a state with meaningful overlap with the ground state you are trying to measure.
QPE returns eigenphase $\varphi_k$ with probability $|\langle\psi_k|\psi_{\text{trial}}\rangle|^2$. If your trial state has 1% overlap with the ground state, you need ~100 runs. If it has $10^{-6}$ overlap — routine for a large molecule — you need a million runs of a circuit that already needs error correction. This case study quantifies the problem, explains why it worsens exponentially with system size, and surveys what is actually done about it.
Skills applied
- Relating QPE success probability to trial-state overlap (§16.9).
- Understanding the orthogonality catastrophe and its scaling.
- Evaluating state-preparation strategies on cost and reliability.
- Assessing end-to-end feasibility rather than the headline subroutine.
Phase 1: The overlap requirement
QPE on a trial state $|\psi_{\text{trial}}\rangle = \sum_k c_k|\psi_k\rangle$ returns eigenphase $k$ with probability $|c_k|^2$. Expected repetitions to observe the ground state:
$$R \approx \frac{1}{|c_0|^2}$$
| Overlap $|c_0|^2$ | Repetitions |
|---|---|
| 0.9 | ~1 |
| 0.5 | 2 |
| 0.1 | 10 |
| 0.01 | 100 |
| $10^{-4}$ | 10,000 |
| $10^{-6}$ | 1,000,000 |
Each repetition is a full QPE circuit — for a chemistry problem, $10^{9}$–$10^{11}$ gates under error correction. A million repetitions multiplies an already enormous circuit by six orders of magnitude.
Phase 2: Why overlap decays exponentially
The standard trial state is the Hartree–Fock determinant, cheap to prepare (a product state, one $X$ gate per occupied orbital). Its overlap with the true ground state is
$$|c_0|^2 \approx \prod_{i=1}^{M} (1 - \epsilon_i)$$
where each of the $M$ correlated orbitals contributes a small loss. For a fixed per-orbital correlation error $\epsilon$, the overlap decays as $(1-\epsilon)^M$ — exponentially in system size.
| Molecule | Orbitals | Typical HF overlap |
|---|---|---|
| H₂ | 4 | ~0.99 |
| LiH | 12 | ~0.95 |
| H₂O | 14 | ~0.90 |
| N₂ (stretched) | 20 | ~0.4 |
| FeMoco (nitrogenase) | ~150 | $\ll 10^{-3}$ |
This is the orthogonality catastrophe: in a large Hilbert space, any fixed simple state becomes nearly orthogonal to the true ground state. And the molecules of scientific interest — strongly correlated transition-metal complexes, stretched bonds, catalytic transition states — are precisely those where Hartree–Fock is worst, because strong correlation is what makes them classically hard in the first place.
The uncomfortable coupling. The systems where quantum advantage is most plausible are the systems where trial-state preparation is hardest. The difficulty is not a coincidence; both stem from the same physics.
Phase 3: What is done about it
1. Better classical trial states. Prepare a CISD, CASSCF, or matrix-product-state ansatz instead of Hartree–Fock. Overlap improves substantially, and preparation costs more gates — an MPS with bond dimension $\chi$ costs roughly $O(n\chi^2)$ gates. This is the most practical current approach and it inherits the limits of the classical method used.
2. Adiabatic state preparation. Start in the ground state of a simple Hamiltonian and evolve slowly to the target. The adiabatic theorem guarantees you stay in the ground state if evolution is slow relative to the inverse square of the minimum spectral gap. For systems with small gaps — again, exactly the strongly correlated ones — the required time is prohibitive, and the gap is generally unknown in advance.
3. Variational preparation. Run VQE (Chapter 19) to obtain an approximate ground state, then use it as the QPE input. This is an appealing hybrid: VQE is shallow enough for near-term hardware and needs only decent overlap, while QPE supplies the precision VQE cannot. The catch is that VQE's own convergence for strongly correlated systems is not guaranteed.
4. Filtering and amplification. Apply spectral filters, or use amplitude amplification to boost the ground-state component. Amplification improves the repetition count from $1/|c_0|^2$ to $1/|c_0|$ — a quadratic saving. For $|c_0|^2 = 10^{-6}$ that is a million repetitions down to a thousand, which is a genuine and substantial improvement.
5. Accept an excited state. Sometimes the relevant quantity is a gap or a reaction barrier rather than an absolute ground-state energy, and a well-characterized excited state suffices.
Phase 4: End-to-end feasibility
Consider FeMoco, the canonical "quantum computers will solve this" example:
| Component | Estimate |
|---|---|
| Logical qubits | ~2,000 |
| QPE circuit (Toffoli) | ~$10^{10}$ |
| HF overlap | $\ll 10^{-3}$ |
| Repetitions without amplification | $> 10^{3}$ |
| Repetitions with amplitude amplification | ~30 |
| Effective total | $\sim3\times10^{11}$ Toffoli |
Amplitude amplification and better trial states are not optional refinements; without them the calculation is out of reach by orders of magnitude even granting the error-corrected hardware.
The honest framing: QPE for chemistry is a fault-tolerant application whose cost is set as much by state preparation as by the eigenvalue estimation itself, and the published resource estimates that omit preparation are not describing an end-to-end algorithm.
Phase 5: How to evaluate a claim
When someone proposes QPE for a chemistry or materials problem, ask:
- What trial state, and what is its estimated overlap? If the answer is "Hartree–Fock" for a strongly correlated system, the estimate is incomplete.
- Is overlap estimated or assumed? Overlap is generally not computable classically for the hard cases — that is the whole problem — so it is usually assumed. Assumed overlap is the single most common gap in resource estimates.
- Does the estimate include repetitions? Multiply the circuit cost by $1/|c_0|^2$, or $1/|c_0|$ with amplification.
- What is the classical baseline? DMRG and coupled-cluster methods are strong; the quantum advantage must be measured against them, not against exact diagonalization.
Discussion Questions
- Overlap decays exponentially with system size while QPE's circuit cost grows polynomially. Which dominates, and at what scale?
- Why are the systems most attractive for quantum simulation also the hardest to prepare states for? Is that avoidable?
- Amplitude amplification improves repetitions from $1/|c_0|^2$ to $1/|c_0|$. Explain the mechanism by analogy to Grover.
- Trial-state overlap usually cannot be computed classically for the interesting cases. What does that imply for the reliability of published resource estimates?
Your Turn: Extensions
- Compute the HF overlap with the exact ground state for H₂ at several bond lengths; observe the collapse at stretched geometries.
- Implement QPE on a 2-qubit Hamiltonian with a deliberately poor trial state and measure the repetitions needed.
- Estimate the gate cost of preparing an MPS trial state with bond dimension 64 on 20 qubits.
- Read a published FeMoco resource estimate and identify its state-preparation assumption.
Key Takeaways
- QPE returns an eigenphase with probability equal to the squared overlap of the trial state, so poor overlap multiplies the cost by $1/|c_0|^2$.
- Hartree–Fock overlap decays exponentially with system size — the orthogonality catastrophe.
- The systems where quantum advantage is most plausible are exactly those where trial-state preparation is hardest; the two difficulties share a cause.
- Amplitude amplification buys a quadratic improvement in repetitions and is essential rather than optional at realistic scales.
- A resource estimate that omits state preparation is not an end-to-end estimate. Ask what trial state, what overlap, and how it was determined.